Obsidian voice notes, transcribed for free

By Sam Ridder — I build and run ScribeGrab on my own hardware, on my own. Who I am.

The Whisper plugins for Obsidian let you choose the server they talk to. Most people leave it on OpenAI and pay $0.006 for every minute they dictate. Change one field to ScribeGrab and the same plugin keeps working — on my graphics card, for nothing, without an API key.

The one field that matters

In the plugin's settings there is an API URL box, filled in with https://api.openai.com/v1/audio/transcriptions. Replace it with:

https://scribegrab.com/v1/audio/transcriptions

That's the whole setup. The endpoint speaks OpenAI's transcription API — same multipart form, same JSON back — so the plugin doesn't know or care that it isn't talking to OpenAI anymore.

Every setting, spelled out

Field names below are from nikdanilov's Whisper plugin, the popular one. Other plugins use the same words in a different order.

Check it in ten seconds

Before you trust it with a real recording, throw any audio file at it from a terminal:

curl https://scribegrab.com/v1/audio/transcriptions \
  -F [email protected] \
  -F model=whisper-1

{"text":"This is the first sentence of the memo."}

Add -F response_format=srt for time-coded subtitles, or verbose_json for a segment list with start and end times. The request stays open until the transcript exists, exactly like OpenAI's — a two-minute voice note usually comes back in a few seconds.

The limits, honestly

This is one person's computer, not a cloud region. Everything below is the same rule the website itself runs on, not a special API tier:

• 90 minutes and 2 GB per file. Longer than that and you get a clear error instead of a hang.
• Two files at a time per person. A third request while two are running gets a 429 with an explanation — queue them one after another.
• You share the queue with the website. When the cards are busy your request waits its turn; when the queue is more than twenty minutes deep the server says so and asks you to come back rather than leaving you hanging. Nothing is lost, you just retry.
• No uptime promise. I restart this machine when it needs it. If a request fails, the plugin shows the error and your audio is still in your vault.
• verbose_json is the plain version. You get id, start, end and text per segment. Whisper's internal statistics (avg_logprob, tokens, no_speech_prob) are not there — I'd rather leave a field out than invent a number.
• No word-level timestamps in the API response, and no speaker labels. Those exist on the website, where you can see and download them.

What happens to your recording

Your audio goes to a machine in my house, gets transcribed, and the uploaded file is deleted the moment the transcript exists. The transcript itself is wiped within 45 minutes. Nothing is kept, sold, or used to train anything, and your recording is never forwarded to OpenAI, Google or any other API — the whole point of this site is that it doesn't pay per minute either. Details in the privacy policy.

That said: it is still my machine. If a recording is confidential enough that a stranger's server is a problem, run Whisper locally instead — that advice costs me nothing and it's the right advice.

Not just Obsidian

Anything that lets you set an OpenAI base URL can use this: dictation launchers, shell scripts, a Makefile that subtitles your screen recordings. If your tool asks for a base URL rather than a full endpoint, give it https://scribegrab.com/v1 and it will append the rest itself.

ScribeGrab is not affiliated with OpenAI or Obsidian; "OpenAI-compatible" here means it accepts the same request and returns the same shape of answer.

Questions

Try it in the browser first →